Can I use openai's whisper instead of faster-whisper with this library? about whisperx

[kirkd7]

I have this same question, so far in my testing original Whisper is more accurate...

[MengHao666]

similary questions. Whisperx even could not return word-level timestamps for the thailand language, it only return the sentence-level timestamps. I am very confused now. Any idea?

[utility-aagrawal]

@MengHao666 , Did you use alignment after transcribing your file? Alignment returns the word-level timestamps but I have only tested it for English. You might want to give this a try for Thai.

audio = whisperx.load_audio(filepath)
result = whisper_model.transcribe(audio)
device = "cuda" if torch.cuda.is_available() else "cpu"
result_aligned = whisperx.align(result["segments"], models["Align_model"], models["Align_metadata"], audio, device, return_char_alignments=False)

[MengHao666]

@MengHao666 , Did you use alignment after transcribing your file? Alignment returns the word-level timestamps but I have only tested it for English. You might want to give this a try for Thai.

audio = whisperx.load_audio(filepath) result = whisper_model.transcribe(audio) device = "cuda" if torch.cuda.is_available() else "cpu" result_aligned = whisperx.align(result["segments"], models["Align_model"], models["Align_metadata"], audio, device, return_char_alignments=False)

I tried, it just failed. It could noit return word-level timestamps. I guess whisperx could not find proper thai language wav-vec2 pretrained model. I have tried to find some in the hugging face. But it failed also.

[wahahaer]

the native whisper is the best for ASR， but without accurate timestamp

[gunwant11]

I just made a loop to format openai/wisper timestamps similar to wisperX output which worked for me, also did string manipulation to append and prepend the punctuations to make it similar to fastwisper output.